Noise covariance model for time-series InSAR analysis Piyush Agram, Mark Simons
Motivation Noise covariance in individual interferograms has been well studied. (Hanssen, 2001) Extend the noise model to networks of interferograms time-series applications. Develop uncertainty measures for time-series InSAR products. Build an InSAR noise covariance model from first principles.
Independent Observations? Simplest interferogram network B A C X Z Y Simple network of coherent pixels in a single interferogram. Is information contained in interferogram BC different from information in AB and AC? Is information at X,Y and Z pixels independent? Conventionally any covariances are ignored. Simplicity. Ease of implementing algorithms.
Single pixel phase model Atmospheric Phase Screen Decorrelation phase noise Phase noise from uncorrelated sources Noise terms Spatially correlated. Correlated in pairs with common scene. Spatially independent. Temporally correlated. Spatially and Temporally independent. Pixel x in interferogram (i,j) i and j are indices of SAR acquisitions.
Vector Phase model M Interferograms, N SAR scenes, P pixels A represents the incidence matrix [+1,0,-1]. Pixel-by-pixel stack of phase observations.
Phase Covariance model In absence of deformation, our noise estimate is Assuming statistical independence of the noise terms. Total covariance matrix is a combination of three models. Atmosphere Scattering / decorrelation System noise
Atmospheric Phase Models S. Jonsson (PhD thesis) Derived from the data itself using radon transform. Hanssen (2001) Emardson et al. (2003) Lohman & Simons (2005)
APS Covariance model 1) Spatial covariance for a SAR scene. 2) Covariance of all pixels in one SAR scene. 3) Covariance of all pixels in all SAR scenes. 4) Covariance of all pixels in all IFGs.
Decorrelation model B Zebker and Villasenor (1992), Bamler and Just (1993). SAR pixel made up of infinite Gaussian scatterers.
Observable: InSAR coherence Just and Bamler (1994) Gaussian signal model InSAR coherence is a function of phase noise. We derive a model to relate coherence directly to noise covariance matrix. Actual temporal behavior of scatterers not known. We use a signal model instead.
Coherence => Signal Model => Stats SAR phase correlation InSAR phase variance Observed InSAR coherence Number of looks taken into account for the mapping.
Decorrelation covariance model 1) SAR phase correlation of one pixel from InSAR coherence. 2) Pseudo-correlation of one pixel in all interferograms. 3) Covariance of one pixel in all IFGs. (Pixels are uncorrelated with each other.) 4) Repeat for every pixel.
System (processor) noise Two SAR scenes A and B. Form two interferograms AB, BA. AB * BA = 0 (theoretically). We observe a noise pattern. Noise levels for ALOS => 0.15 rads or 8 degs. Estimates using Stanford mocomp processor over Parkfield, CA. Diagonal structure.
Spatial structure: Implication SBAS => pixel-by-pixel inversion. MInTS => wavelet coefficient-by-coefficient. Poster Today: Multi-scale InSAR Time Series (MInTS) analysis of the creeping section of the San Andreas Fault
What does MInTS do? C d = A exp(l/l c ) L c = 25 km Corresponding wavelets coefficients nearly uncorrelated
Temporal correlation: Implication Hanssen (2001) is the only correlated noise component in a network of IFGs. => diagonal covariance matrix. Our model Consistent with decorrelation models. Even if the atmospheric phase signal could be perfectly corrected, scatterer noise would still be correlated in time. Weighting scheme for InSAR observations in time domain. (Not in MInTS yet)
Thus far We have a presented a covariance model in space and time Accounts for System noise Interferograms with common scene Decorrelation phase model Atmospheric propagation delays Does not yet account for Orbit errors Tropostatic delays Ionospheric delays
Application: Optimal IFG networks Questions of Interest for data centers 1. Given M SAR scenes, what is the optimal IFG network of size N? 2. Given a network of IFGs, how do I augment it optimally with new SAR scenes? is the temporal baseline matrix. is the LOS velocity. C is the uncertainty in estimated velocity.
Algorithms: Subset of Observations M SAR scenes = 2^(M 2 /2) IFG networks. Impossible to evaluate all networks. Subset of observations (Reeves and Zhe, 1999). Sequential Backward Selection (SBS) Remove IFG that contributes the least at each step. Sequential Hybrid Selection (SHS) Allows IFG already removed to re-enter optimal solution at later stage. Any reasonable metric chosen to optimize network we choose LOS velocity uncertainty for this work.
Example: 9 SAR scenes Monthly acquisitions with zero geometric baseline. Decorrelation time constant of 0.36 yr and 1 cm atmosphere. Total number of viable IFGS = 21 8 one hop IFGs (blue) 7 two hop IFGs (green) 6 three hop IFGs (red)
Example: Pruning networks Optimal 15/21 IFG network using SBS/ SHS. Uncertainty ratio = 1.05 Optimal 8/21 IFG network using SBS/ SHS. Uncertainty ratio = 1.12
Network design 9 SAR scenes 30 SAR scenes 1. SBS/ SHS allows us to design networks of any size. 2. Allows us to prioritize processing of IFGs. 3. Great for designing networks with large number of acquisitions.
Example: Augmenting networks Adding 1 new SAR scenes and 2 IFGs to existing networks.
Conclusions InSAR phase observations are correlated over space and time - not independent observations. Accounting for spatial correlation significantly improves deformation estimates (e.g. MInTS). Observed InSAR coherence can be directly translated to a temporal covariance matrix using a model. A noise covariance model allows us to provide conservative error estimates with time-series products.